Toward robust multimodal fake news detection under weak supervision via text-derived surrogate labels

Fake news dissemination on social media represents a major societal challenge, affecting public trust, democratic processes, and the reliability of online information ecosystems. Recent multimodal approaches have shown that jointly exploiting textual content, visual information, and social propagation patterns substantially improves fake news detection performance. However, most existing methods rely on fully supervised settings requiring large amounts of manually verified annotations, an assumption that rarely holds in real-world environments where reliable labels are scarce and costly to obtain. To address this limitation, we propose M3DUSA-WS, a weakly supervised extension of the M3DUSA framework for robust fake news detection under limited ground-truth availability. Motivated by the strong predictive performance and robustness previously demonstrated by the fully supervised approach, the proposed framework extends its early-fusion heterogeneous graph formulation by incorporating weak supervision through text-derived surrogate labels. All available information is represented within a unified heterogeneous information network, enabling the joint modeling of multimodal content and relational structure through a Graph Neural Network encoder. The learned graph representations are processed by two complementary prediction heads: a main classifier guided by verified labels and an auxiliary classifier relying on surrogate labels generated by a ROBERTA-based model trained on external fake news datasets. A combined optimization objective integrates supervised classification, surrogate supervision, and consistency regularization, encouraging agreement between the two prediction heads while mitigating the impact of noisy weak labels.